COMPAS

The investigation found that Black defendants were nearly twice as likely to be incorrectly classified as high-risk compared to white defendants. Specifically, the false positive rate for Black defendants was 44.9%, while for white defendants it was 23.5%. This means the algorithm labeled Black individuals as future violent criminals when they did not reoffend at almost double the rate.

Conversely, the algorithm systematically underestimated risk for white defendants. White offenders were more likely to be incorrectly labeled as low-risk and subsequently reoffend. The false negative rate for white defendants was 47.7%, compared to 28.0% for Black defendants. This dual asymmetry—higher false positives for Black individuals and higher false negatives for white individuals—demonstrates a structural failure in equalized odds.

Northpointe, the creator of COMPAS, defended the algorithm by arguing it achieved accuracy parity: the overall accuracy rate was roughly equal across racial groups at approximately 65%. However, ProPublica demonstrated that overall accuracy masks critical differences in error distribution. The investigation highlighted the tension between predictive parity (equal positive predictive value across groups) and error rate balance, a debate that continues to shape fairness metric selection today.

The investigation scrutinized the 137-question survey underlying COMPAS. Several questions were found to be proxies for socioeconomic status and race rather than direct measures of criminal propensity. Examples include: 'Was one of your parents ever sent to jail or prison?' and 'How often have you moved in the last twelve months?' These questions encode historical bias, reflecting systemic inequalities in policing and incarceration rather than individual risk.

COMPAS scores were presented to judges as a 1-to-10 risk scale during pre-trial and sentencing decisions. The investigation revealed that judges often relied heavily on these scores, despite the tool's proprietary nature preventing independent validation. This case became a landmark example of automation bias in high-stakes contexts, where decision-makers over-delegate judgment to algorithmic outputs without understanding their limitations.

ProPublica obtained risk scores for over 7,000 individuals arrested in Broward County, Florida, between 2013 and 2014. The analysis tracked whether each individual was charged with a new crime over the next two years. This methodology established a replicable framework for bias audits using publicly available criminal records. The full dataset and methodology were published openly, setting a precedent for algorithmic transparency in investigative journalism.

The study and practice of designing machine learning systems that make decisions without unjustified discrimination. The COMPAS case is a canonical example because an investigation by ProPublica found the tool exhibited racial bias: it was twice as likely to falsely flag Black defendants as future criminals compared to white defendants, while making the opposite error for white defendants at a higher rate.

A fairness criterion requiring a classifier to achieve equal true positive rates and equal false positive rates across protected groups. The COMPAS controversy highlighted a fundamental impossibility: the tool satisfied predictive parity (equal precision across groups) but violated equalized odds, proving that multiple fairness definitions cannot be simultaneously satisfied in practice.

A legal doctrine identifying facially neutral policies that disproportionately harm protected classes. Under the U.S. Four-Fifths Rule, a selection rate for any group less than 80% of the highest group's rate constitutes evidence of adverse impact. COMPAS's risk scores, while not using race as an explicit input, produced outcomes that would trigger disparate impact scrutiny under this framework.

The observed tension where enforcing strict fairness constraints can reduce overall predictive accuracy. In COMPAS, the tool's designers argued for predictive parity—ensuring a risk score of 7 meant the same likelihood of recidivism regardless of race. Critics countered that this came at the cost of equalized odds, forcing a trade-off between two mathematically incompatible fairness definitions.

A causal definition of fairness where a decision is fair if it would remain the same in a counterfactual world where the individual belonged to a different demographic group. Applied to COMPAS, this framework would ask: would the defendant's risk score change if their race were different, holding all causally relevant factors constant? This requires explicit causal modeling of how race influences input features like prior arrests.